i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         4
  25 #define NB_ASM_REGS     8
  26
  27 /* a register can belong to several classes. The classes must be
  28    sorted from more general to more precise (see gv2() code which does
  29    assumptions on it). */
  30 #define RC_INT     0x0001 /* generic integer register */
  31 #define RC_FLOAT   0x0002 /* generic float register */
  32 #define RC_EAX     0x0004
  33 #define RC_ST0     0x0008
  34 #define RC_ECX     0x0010
  35 #define RC_EDX     0x0020
  36 #define RC_IRET    RC_EAX /* function return: integer register */
  37 #define RC_LRET    RC_EDX /* function return: second integer register */
  38 #define RC_FRET    RC_ST0 /* function return: float register */
  39
  40 /* pretty names for the registers */
  41 enum {
  42     TREG_EAX = 0,
  43     TREG_ECX,
  44     TREG_EDX,
  45     TREG_ST0,
  46     TREG_ESP = 4
  47 };
  48
  49 /* return registers for function */
  50 #define REG_IRET TREG_EAX /* single word int return register */
  51 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  52 #define REG_FRET TREG_ST0 /* float return register */
  53
  54 /* defined if function parameters must be evaluated in reverse order */
  55 #define INVERT_FUNC_PARAMS
  56
  57 /* defined if structures are passed as pointers. Otherwise structures
  58    are directly pushed on stack. */
  59 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  60
  61 /* pointer size, in bytes */
  62 #define PTR_SIZE 4
  63
  64 /* long double size and alignment, in bytes */
  65 #define LDOUBLE_SIZE  12
  66 #define LDOUBLE_ALIGN 4
  67 /* maximum alignment (for aligned attribute support) */
  68 #define MAX_ALIGN     8
  69
  70
  71 #define psym oad
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_DATA_PTR  R_386_32
  81 #define R_JMP_SLOT  R_386_JMP_SLOT
  82 #define R_COPY      R_386_COPY
  83
  84 #define ELF_START_ADDR 0x08048000
  85 #define ELF_PAGE_SIZE  0x1000
  86
  87 /******************************************************/
  88 #else /* ! TARGET_DEFS_ONLY */
  89 /******************************************************/
  90 #include "tcc.h"
  91
  92 ST_DATA const int reg_classes[NB_REGS] = {
  93     /* eax */ RC_INT | RC_EAX,
  94     /* ecx */ RC_INT | RC_ECX,
  95     /* edx */ RC_INT | RC_EDX,
  96     /* st0 */ RC_FLOAT | RC_ST0,
  97 };
  98
  99 static unsigned long func_sub_sp_offset;
 100 static int func_ret_sub;
 101 #ifdef CONFIG_TCC_BCHECK
 102 static addr_t func_bound_offset;
 103 #endif
 104
 105 /* XXX: make it faster ? */
 106 ST_FUNC void g(int c)
 107 {
 108     int ind1;
 109     ind1 = ind + 1;
 110     if (ind1 > cur_text_section->data_allocated)
 111         section_realloc(cur_text_section, ind1);
 112     cur_text_section->data[ind] = c;
 113     ind = ind1;
 114 }
 115
 116 ST_FUNC void o(unsigned int c)
 117 {
 118     while (c) {
 119         g(c);
 120         c = c >> 8;
 121     }
 122 }
 123
 124 ST_FUNC void gen_le16(int v)
 125 {
 126     g(v);
 127     g(v >> 8);
 128 }
 129
 130 ST_FUNC void gen_le32(int c)
 131 {
 132     g(c);
 133     g(c >> 8);
 134     g(c >> 16);
 135     g(c >> 24);
 136 }
 137
 138 /* output a symbol and patch all calls to it */
 139 ST_FUNC void gsym_addr(int t, int a)
 140 {
 141     int n, *ptr;
 142     while (t) {
 143         ptr = (int *)(cur_text_section->data + t);
 144         n = *ptr; /* next value */
 145         *ptr = a - t - 4;
 146         t = n;
 147     }
 148 }
 149
 150 ST_FUNC void gsym(int t)
 151 {
 152     gsym_addr(t, ind);
 153 }
 154
 155 /* psym is used to put an instruction with a data field which is a
 156    reference to a symbol. It is in fact the same as oad ! */
 157 #define psym oad
 158
 159 /* instruction + 4 bytes data. Return the address of the data */
 160 ST_FUNC int oad(int c, int s)
 161 {
 162     int ind1;
 163
 164     o(c);
 165     ind1 = ind + 4;
 166     if (ind1 > cur_text_section->data_allocated)
 167         section_realloc(cur_text_section, ind1);
 168     *(int *)(cur_text_section->data + ind) = s;
 169     s = ind;
 170     ind = ind1;
 171     return s;
 172 }
 173
 174 /* output constant with relocation if 'r & VT_SYM' is true */
 175 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 176 {
 177     if (r & VT_SYM)
 178         greloc(cur_text_section, sym, ind, R_386_32);
 179     gen_le32(c);
 180 }
 181
 182 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 183 {
 184     if (r & VT_SYM)
 185         greloc(cur_text_section, sym, ind, R_386_PC32);
 186     gen_le32(c - 4);
 187 }
 188
 189 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 190    opcode bits */
 191 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 192 {
 193     op_reg = op_reg << 3;
 194     if ((r & VT_VALMASK) == VT_CONST) {
 195         /* constant memory reference */
 196         o(0x05 | op_reg);
 197         gen_addr32(r, sym, c);
 198     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 199         /* currently, we use only ebp as base */
 200         if (c == (char)c) {
 201             /* short reference */
 202             o(0x45 | op_reg);
 203             g(c);
 204         } else {
 205             oad(0x85 | op_reg, c);
 206         }
 207     } else {
 208         g(0x00 | op_reg | (r & VT_VALMASK));
 209     }
 210 }
 211
 212 /* load 'r' from value 'sv' */
 213 ST_FUNC void load(int r, SValue *sv)
 214 {
 215     int v, t, ft, fc, fr;
 216     SValue v1;
 217
 218 #ifdef TCC_TARGET_PE
 219     SValue v2;
 220     sv = pe_getimport(sv, &v2);
 221 #endif
 222
 223     fr = sv->r;
 224     ft = sv->type.t;
 225     fc = sv->c.ul;
 226
 227     v = fr & VT_VALMASK;
 228     if (fr & VT_LVAL) {
 229         if (v == VT_LLOCAL) {
 230             v1.type.t = VT_INT;
 231             v1.r = VT_LOCAL | VT_LVAL;
 232             v1.c.ul = fc;
 233             fr = r;
 234             if (!(reg_classes[fr] & RC_INT))
 235                 fr = get_reg(RC_INT);
 236             load(fr, &v1);
 237         }
 238         if ((ft & VT_BTYPE) == VT_FLOAT) {
 239             o(0xd9); /* flds */
 240             r = 0;
 241         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 242             o(0xdd); /* fldl */
 243             r = 0;
 244         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 245             o(0xdb); /* fldt */
 246             r = 5;
 247         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 248             o(0xbe0f);   /* movsbl */
 249         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 250             o(0xb60f);   /* movzbl */
 251         } else if ((ft & VT_TYPE) == VT_SHORT) {
 252             o(0xbf0f);   /* movswl */
 253         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 254             o(0xb70f);   /* movzwl */
 255         } else {
 256             o(0x8b);     /* movl */
 257         }
 258         gen_modrm(r, fr, sv->sym, fc);
 259     } else {
 260         if (v == VT_CONST) {
 261             o(0xb8 + r); /* mov $xx, r */
 262             gen_addr32(fr, sv->sym, fc);
 263         } else if (v == VT_LOCAL) {
 264             if (fc) {
 265                 o(0x8d); /* lea xxx(%ebp), r */
 266                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 267             } else {
 268                 o(0x89);
 269                 o(0xe8 + r); /* mov %ebp, r */
 270             }
 271         } else if (v == VT_CMP) {
 272             oad(0xb8 + r, 0); /* mov $0, r */
 273             o(0x0f); /* setxx %br */
 274             o(fc);
 275             o(0xc0 + r);
 276         } else if (v == VT_JMP || v == VT_JMPI) {
 277             t = v & 1;
 278             oad(0xb8 + r, t); /* mov $1, r */
 279             o(0x05eb); /* jmp after */
 280             gsym(fc);
 281             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 282         } else if (v != r) {
 283             o(0x89);
 284             o(0xc0 + r + v * 8); /* mov v, r */
 285         }
 286     }
 287 }
 288
 289 /* store register 'r' in lvalue 'v' */
 290 ST_FUNC void store(int r, SValue *v)
 291 {
 292     int fr, bt, ft, fc;
 293
 294 #ifdef TCC_TARGET_PE
 295     SValue v2;
 296     v = pe_getimport(v, &v2);
 297 #endif
 298
 299     ft = v->type.t;
 300     fc = v->c.ul;
 301     fr = v->r & VT_VALMASK;
 302     bt = ft & VT_BTYPE;
 303     /* XXX: incorrect if float reg to reg */
 304     if (bt == VT_FLOAT) {
 305         o(0xd9); /* fsts */
 306         r = 2;
 307     } else if (bt == VT_DOUBLE) {
 308         o(0xdd); /* fstpl */
 309         r = 2;
 310     } else if (bt == VT_LDOUBLE) {
 311         o(0xc0d9); /* fld %st(0) */
 312         o(0xdb); /* fstpt */
 313         r = 7;
 314     } else {
 315         if (bt == VT_SHORT)
 316             o(0x66);
 317         if (bt == VT_BYTE || bt == VT_BOOL)
 318             o(0x88);
 319         else
 320             o(0x89);
 321     }
 322     if (fr == VT_CONST ||
 323         fr == VT_LOCAL ||
 324         (v->r & VT_LVAL)) {
 325         gen_modrm(r, v->r, v->sym, fc);
 326     } else if (fr != r) {
 327         o(0xc0 + fr + r * 8); /* mov r, fr */
 328     }
 329 }
 330
 331 static void gadd_sp(int val)
 332 {
 333     if (val == (char)val) {
 334         o(0xc483);
 335         g(val);
 336     } else {
 337         oad(0xc481, val); /* add $xxx, %esp */
 338     }
 339 }
 340
 341 static void gen_static_call(int v)
 342 {
 343     Sym *sym;
 344
 345     sym = external_global_sym(v, &func_old_type, 0);
 346     oad(0xe8, -4);
 347     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 348 }
 349
 350 /* 'is_jmp' is '1' if it is a jump */
 351 static void gcall_or_jmp(int is_jmp)
 352 {
 353     int r;
 354     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 355         /* constant case */
 356         if (vtop->r & VT_SYM) {
 357             /* relocation case */
 358             greloc(cur_text_section, vtop->sym,
 359                    ind + 1, R_386_PC32);
 360         } else {
 361             /* put an empty PC32 relocation */
 362             put_elf_reloc(symtab_section, cur_text_section,
 363                           ind + 1, R_386_PC32, 0);
 364         }
 365         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 366     } else {
 367         /* otherwise, indirect call */
 368         r = gv(RC_INT);
 369         o(0xff); /* call/jmp *r */
 370         o(0xd0 + r + (is_jmp << 4));
 371     }
 372 }
 373
 374 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 375 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 376
 377 /* Return the number of registers needed to return the struct, or 0 if
 378    returning via struct pointer. */
 379 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 380 {
 381 #ifdef TCC_TARGET_PE
 382     int size, align;
 383
 384     *ret_align = 1; // Never have to re-align return values for x86
 385     *regsize = 4;
 386     size = type_size(vt, &align);
 387     if (size > 8) {
 388         return 0;
 389     } else if (size > 4) {
 390         ret->ref = NULL;
 391         ret->t = VT_LLONG;
 392         return 1;
 393     } else {
 394         ret->ref = NULL;
 395         ret->t = VT_INT;
 396         return 1;
 397     }
 398 #else
 399     *ret_align = 1; // Never have to re-align return values for x86
 400     return 0;
 401 #endif
 402 }
 403
 404 /* Generate function call. The function address is pushed first, then
 405    all the parameters in call order. This functions pops all the
 406    parameters and the function address. */
 407 ST_FUNC void gfunc_call(int nb_args)
 408 {
 409     int size, align, r, args_size, i, func_call;
 410     Sym *func_sym;
 411
 412     args_size = 0;
 413     for(i = 0;i < nb_args; i++) {
 414         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 415             size = type_size(&vtop->type, &align);
 416             /* align to stack align size */
 417             size = (size + 3) & ~3;
 418             /* allocate the necessary size on stack */
 419             oad(0xec81, size); /* sub $xxx, %esp */
 420             /* generate structure store */
 421             r = get_reg(RC_INT);
 422             o(0x89); /* mov %esp, r */
 423             o(0xe0 + r);
 424             vset(&vtop->type, r | VT_LVAL, 0);
 425             vswap();
 426             vstore();
 427             args_size += size;
 428         } else if (is_float(vtop->type.t)) {
 429             gv(RC_FLOAT); /* only one float register */
 430             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 431                 size = 4;
 432             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 433                 size = 8;
 434             else
 435                 size = 12;
 436             oad(0xec81, size); /* sub $xxx, %esp */
 437             if (size == 12)
 438                 o(0x7cdb);
 439             else
 440                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 441             g(0x24);
 442             g(0x00);
 443             args_size += size;
 444         } else {
 445             /* simple type (currently always same size) */
 446             /* XXX: implicit cast ? */
 447             r = gv(RC_INT);
 448             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 449                 size = 8;
 450                 o(0x50 + vtop->r2); /* push r */
 451             } else {
 452                 size = 4;
 453             }
 454             o(0x50 + r); /* push r */
 455             args_size += size;
 456         }
 457         vtop--;
 458     }
 459     save_regs(0); /* save used temporary registers */
 460     func_sym = vtop->type.ref;
 461     func_call = func_sym->a.func_call;
 462     /* fast call case */
 463     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 464         func_call == FUNC_FASTCALLW) {
 465         int fastcall_nb_regs;
 466         uint8_t *fastcall_regs_ptr;
 467         if (func_call == FUNC_FASTCALLW) {
 468             fastcall_regs_ptr = fastcallw_regs;
 469             fastcall_nb_regs = 2;
 470         } else {
 471             fastcall_regs_ptr = fastcall_regs;
 472             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 473         }
 474         for(i = 0;i < fastcall_nb_regs; i++) {
 475             if (args_size <= 0)
 476                 break;
 477             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 478             /* XXX: incorrect for struct/floats */
 479             args_size -= 4;
 480         }
 481     }
 482 #ifndef TCC_TARGET_PE
 483     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 484         args_size -= 4;
 485 #endif
 486     gcall_or_jmp(0);
 487
 488     if (args_size && func_call != FUNC_STDCALL)
 489         gadd_sp(args_size);
 490     vtop--;
 491 }
 492
 493 #ifdef TCC_TARGET_PE
 494 #define FUNC_PROLOG_SIZE 10
 495 #else
 496 #define FUNC_PROLOG_SIZE 9
 497 #endif
 498
 499 /* generate function prolog of type 't' */
 500 ST_FUNC void gfunc_prolog(CType *func_type)
 501 {
 502     int addr, align, size, func_call, fastcall_nb_regs;
 503     int param_index, param_addr;
 504     uint8_t *fastcall_regs_ptr;
 505     Sym *sym;
 506     CType *type;
 507
 508     sym = func_type->ref;
 509     func_call = sym->a.func_call;
 510     addr = 8;
 511     loc = 0;
 512     func_vc = 0;
 513
 514     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 515         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 516         fastcall_regs_ptr = fastcall_regs;
 517     } else if (func_call == FUNC_FASTCALLW) {
 518         fastcall_nb_regs = 2;
 519         fastcall_regs_ptr = fastcallw_regs;
 520     } else {
 521         fastcall_nb_regs = 0;
 522         fastcall_regs_ptr = NULL;
 523     }
 524     param_index = 0;
 525
 526     ind += FUNC_PROLOG_SIZE;
 527     func_sub_sp_offset = ind;
 528     /* if the function returns a structure, then add an
 529        implicit pointer parameter */
 530     func_vt = sym->type;
 531     func_var = (sym->c == FUNC_ELLIPSIS);
 532 #ifdef TCC_TARGET_PE
 533     size = type_size(&func_vt,&align);
 534     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 535 #else
 536     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 537 #endif
 538         /* XXX: fastcall case ? */
 539         func_vc = addr;
 540         addr += 4;
 541         param_index++;
 542     }
 543     /* define parameters */
 544     while ((sym = sym->next) != NULL) {
 545         type = &sym->type;
 546         size = type_size(type, &align);
 547         size = (size + 3) & ~3;
 548 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 549         /* structs are passed as pointer */
 550         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 551             size = 4;
 552         }
 553 #endif
 554         if (param_index < fastcall_nb_regs) {
 555             /* save FASTCALL register */
 556             loc -= 4;
 557             o(0x89);     /* movl */
 558             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 559             param_addr = loc;
 560         } else {
 561             param_addr = addr;
 562             addr += size;
 563         }
 564         sym_push(sym->v & ~SYM_FIELD, type,
 565                  VT_LOCAL | lvalue_type(type->t), param_addr);
 566         param_index++;
 567     }
 568     func_ret_sub = 0;
 569     /* pascal type call ? */
 570     if (func_call == FUNC_STDCALL)
 571         func_ret_sub = addr - 8;
 572 #ifndef TCC_TARGET_PE
 573     else if (func_vc)
 574         func_ret_sub = 4;
 575 #endif
 576
 577 #ifdef CONFIG_TCC_BCHECK
 578     /* leave some room for bound checking code */
 579     if (tcc_state->do_bounds_check) {
 580         oad(0xb8, 0); /* lbound section pointer */
 581         oad(0xb8, 0); /* call to function */
 582         func_bound_offset = lbounds_section->data_offset;
 583     }
 584 #endif
 585 }
 586
 587 /* generate function epilog */
 588 ST_FUNC void gfunc_epilog(void)
 589 {
 590     addr_t v, saved_ind;
 591
 592 #ifdef CONFIG_TCC_BCHECK
 593     if (tcc_state->do_bounds_check
 594      && func_bound_offset != lbounds_section->data_offset) {
 595         addr_t saved_ind;
 596         addr_t *bounds_ptr;
 597         Sym *sym_data;
 598         /* add end of table info */
 599         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 600         *bounds_ptr = 0;
 601         /* generate bound local allocation */
 602         saved_ind = ind;
 603         ind = func_sub_sp_offset;
 604         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 605                                func_bound_offset, lbounds_section->data_offset);
 606         greloc(cur_text_section, sym_data,
 607                ind + 1, R_386_32);
 608         oad(0xb8, 0); /* mov %eax, xxx */
 609         gen_static_call(TOK___bound_local_new);
 610
 611         ind = saved_ind;
 612         /* generate bound check local freeing */
 613         o(0x5250); /* save returned value, if any */
 614         greloc(cur_text_section, sym_data,
 615                ind + 1, R_386_32);
 616         oad(0xb8, 0); /* mov %eax, xxx */
 617         gen_static_call(TOK___bound_local_delete);
 618
 619         o(0x585a); /* restore returned value, if any */
 620     }
 621 #endif
 622     o(0xc9); /* leave */
 623     if (func_ret_sub == 0) {
 624         o(0xc3); /* ret */
 625     } else {
 626         o(0xc2); /* ret n */
 627         g(func_ret_sub);
 628         g(func_ret_sub >> 8);
 629     }
 630     /* align local size to word & save local variables */
 631
 632     v = (-loc + 3) & -4;
 633     saved_ind = ind;
 634     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 635 #ifdef TCC_TARGET_PE
 636     if (v >= 4096) {
 637         oad(0xb8, v); /* mov stacksize, %eax */
 638         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 639     } else
 640 #endif
 641     {
 642         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 643         o(0xec81);  /* sub esp, stacksize */
 644         gen_le32(v);
 645 #if FUNC_PROLOG_SIZE == 10
 646         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 647 #endif
 648     }
 649     ind = saved_ind;
 650 }
 651
 652 /* generate a jump to a label */
 653 ST_FUNC int gjmp(int t)
 654 {
 655     return psym(0xe9, t);
 656 }
 657
 658 /* generate a jump to a fixed address */
 659 ST_FUNC void gjmp_addr(int a)
 660 {
 661     int r;
 662     r = a - ind - 2;
 663     if (r == (char)r) {
 664         g(0xeb);
 665         g(r);
 666     } else {
 667         oad(0xe9, a - ind - 5);
 668     }
 669 }
 670
 671 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 672 ST_FUNC int gtst(int inv, int t)
 673 {
 674     int v, *p;
 675
 676     v = vtop->r & VT_VALMASK;
 677     if (v == VT_CMP) {
 678         /* fast case : can jump directly since flags are set */
 679         g(0x0f);
 680         t = psym((vtop->c.i - 16) ^ inv, t);
 681     } else if (v == VT_JMP || v == VT_JMPI) {
 682         /* && or || optimization */
 683         if ((v & 1) == inv) {
 684             /* insert vtop->c jump list in t */
 685             p = &vtop->c.i;
 686             while (*p != 0)
 687                 p = (int *)(cur_text_section->data + *p);
 688             *p = t;
 689             t = vtop->c.i;
 690         } else {
 691             t = gjmp(t);
 692             gsym(vtop->c.i);
 693         }
 694     }
 695     vtop--;
 696     return t;
 697 }
 698
 699 /* generate an integer binary operation */
 700 ST_FUNC void gen_opi(int op)
 701 {
 702     int r, fr, opc, c;
 703
 704     switch(op) {
 705     case '+':
 706     case TOK_ADDC1: /* add with carry generation */
 707         opc = 0;
 708     gen_op8:
 709         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 710             /* constant case */
 711             vswap();
 712             r = gv(RC_INT);
 713             vswap();
 714             c = vtop->c.i;
 715             if (c == (char)c) {
 716                 /* generate inc and dec for smaller code */
 717                 if (c==1 && opc==0) {
 718                     o (0x40 | r); // inc
 719                 } else if (c==1 && opc==5) {
 720                     o (0x48 | r); // dec
 721                 } else {
 722                     o(0x83);
 723                     o(0xc0 | (opc << 3) | r);
 724                     g(c);
 725                 }
 726             } else {
 727                 o(0x81);
 728                 oad(0xc0 | (opc << 3) | r, c);
 729             }
 730         } else {
 731             gv2(RC_INT, RC_INT);
 732             r = vtop[-1].r;
 733             fr = vtop[0].r;
 734             o((opc << 3) | 0x01);
 735             o(0xc0 + r + fr * 8);
 736         }
 737         vtop--;
 738         if (op >= TOK_ULT && op <= TOK_GT) {
 739             vtop->r = VT_CMP;
 740             vtop->c.i = op;
 741         }
 742         break;
 743     case '-':
 744     case TOK_SUBC1: /* sub with carry generation */
 745         opc = 5;
 746         goto gen_op8;
 747     case TOK_ADDC2: /* add with carry use */
 748         opc = 2;
 749         goto gen_op8;
 750     case TOK_SUBC2: /* sub with carry use */
 751         opc = 3;
 752         goto gen_op8;
 753     case '&':
 754         opc = 4;
 755         goto gen_op8;
 756     case '^':
 757         opc = 6;
 758         goto gen_op8;
 759     case '|':
 760         opc = 1;
 761         goto gen_op8;
 762     case '*':
 763         gv2(RC_INT, RC_INT);
 764         r = vtop[-1].r;
 765         fr = vtop[0].r;
 766         vtop--;
 767         o(0xaf0f); /* imul fr, r */
 768         o(0xc0 + fr + r * 8);
 769         break;
 770     case TOK_SHL:
 771         opc = 4;
 772         goto gen_shift;
 773     case TOK_SHR:
 774         opc = 5;
 775         goto gen_shift;
 776     case TOK_SAR:
 777         opc = 7;
 778     gen_shift:
 779         opc = 0xc0 | (opc << 3);
 780         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 781             /* constant case */
 782             vswap();
 783             r = gv(RC_INT);
 784             vswap();
 785             c = vtop->c.i & 0x1f;
 786             o(0xc1); /* shl/shr/sar $xxx, r */
 787             o(opc | r);
 788             g(c);
 789         } else {
 790             /* we generate the shift in ecx */
 791             gv2(RC_INT, RC_ECX);
 792             r = vtop[-1].r;
 793             o(0xd3); /* shl/shr/sar %cl, r */
 794             o(opc | r);
 795         }
 796         vtop--;
 797         break;
 798     case '/':
 799     case TOK_UDIV:
 800     case TOK_PDIV:
 801     case '%':
 802     case TOK_UMOD:
 803     case TOK_UMULL:
 804         /* first operand must be in eax */
 805         /* XXX: need better constraint for second operand */
 806         gv2(RC_EAX, RC_ECX);
 807         r = vtop[-1].r;
 808         fr = vtop[0].r;
 809         vtop--;
 810         save_reg(TREG_EDX);
 811         if (op == TOK_UMULL) {
 812             o(0xf7); /* mul fr */
 813             o(0xe0 + fr);
 814             vtop->r2 = TREG_EDX;
 815             r = TREG_EAX;
 816         } else {
 817             if (op == TOK_UDIV || op == TOK_UMOD) {
 818                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 819                 o(0xf0 + fr);
 820             } else {
 821                 o(0xf799); /* cltd, idiv fr, %eax */
 822                 o(0xf8 + fr);
 823             }
 824             if (op == '%' || op == TOK_UMOD)
 825                 r = TREG_EDX;
 826             else
 827                 r = TREG_EAX;
 828         }
 829         vtop->r = r;
 830         break;
 831     default:
 832         opc = 7;
 833         goto gen_op8;
 834     }
 835 }
 836
 837 /* generate a floating point operation 'v = t1 op t2' instruction. The
 838    two operands are guaranted to have the same floating point type */
 839 /* XXX: need to use ST1 too */
 840 ST_FUNC void gen_opf(int op)
 841 {
 842     int a, ft, fc, swapped, r;
 843
 844     /* convert constants to memory references */
 845     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 846         vswap();
 847         gv(RC_FLOAT);
 848         vswap();
 849     }
 850     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 851         gv(RC_FLOAT);
 852
 853     /* must put at least one value in the floating point register */
 854     if ((vtop[-1].r & VT_LVAL) &&
 855         (vtop[0].r & VT_LVAL)) {
 856         vswap();
 857         gv(RC_FLOAT);
 858         vswap();
 859     }
 860     swapped = 0;
 861     /* swap the stack if needed so that t1 is the register and t2 is
 862        the memory reference */
 863     if (vtop[-1].r & VT_LVAL) {
 864         vswap();
 865         swapped = 1;
 866     }
 867     if (op >= TOK_ULT && op <= TOK_GT) {
 868         /* load on stack second operand */
 869         load(TREG_ST0, vtop);
 870         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 871         if (op == TOK_GE || op == TOK_GT)
 872             swapped = !swapped;
 873         else if (op == TOK_EQ || op == TOK_NE)
 874             swapped = 0;
 875         if (swapped)
 876             o(0xc9d9); /* fxch %st(1) */
 877         if (op == TOK_EQ || op == TOK_NE)
 878             o(0xe9da); /* fucompp */
 879         else
 880             o(0xd9de); /* fcompp */
 881         o(0xe0df); /* fnstsw %ax */
 882         if (op == TOK_EQ) {
 883             o(0x45e480); /* and $0x45, %ah */
 884             o(0x40fC80); /* cmp $0x40, %ah */
 885         } else if (op == TOK_NE) {
 886             o(0x45e480); /* and $0x45, %ah */
 887             o(0x40f480); /* xor $0x40, %ah */
 888             op = TOK_NE;
 889         } else if (op == TOK_GE || op == TOK_LE) {
 890             o(0x05c4f6); /* test $0x05, %ah */
 891             op = TOK_EQ;
 892         } else {
 893             o(0x45c4f6); /* test $0x45, %ah */
 894             op = TOK_EQ;
 895         }
 896         vtop--;
 897         vtop->r = VT_CMP;
 898         vtop->c.i = op;
 899     } else {
 900         /* no memory reference possible for long double operations */
 901         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 902             load(TREG_ST0, vtop);
 903             swapped = !swapped;
 904         }
 905
 906         switch(op) {
 907         default:
 908         case '+':
 909             a = 0;
 910             break;
 911         case '-':
 912             a = 4;
 913             if (swapped)
 914                 a++;
 915             break;
 916         case '*':
 917             a = 1;
 918             break;
 919         case '/':
 920             a = 6;
 921             if (swapped)
 922                 a++;
 923             break;
 924         }
 925         ft = vtop->type.t;
 926         fc = vtop->c.ul;
 927         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 928             o(0xde); /* fxxxp %st, %st(1) */
 929             o(0xc1 + (a << 3));
 930         } else {
 931             /* if saved lvalue, then we must reload it */
 932             r = vtop->r;
 933             if ((r & VT_VALMASK) == VT_LLOCAL) {
 934                 SValue v1;
 935                 r = get_reg(RC_INT);
 936                 v1.type.t = VT_INT;
 937                 v1.r = VT_LOCAL | VT_LVAL;
 938                 v1.c.ul = fc;
 939                 load(r, &v1);
 940                 fc = 0;
 941             }
 942
 943             if ((ft & VT_BTYPE) == VT_DOUBLE)
 944                 o(0xdc);
 945             else
 946                 o(0xd8);
 947             gen_modrm(a, r, vtop->sym, fc);
 948         }
 949         vtop--;
 950     }
 951 }
 952
 953 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 954    and 'long long' cases. */
 955 ST_FUNC void gen_cvt_itof(int t)
 956 {
 957     save_reg(TREG_ST0);
 958     gv(RC_INT);
 959     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 960         /* signed long long to float/double/long double (unsigned case
 961            is handled generically) */
 962         o(0x50 + vtop->r2); /* push r2 */
 963         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 964         o(0x242cdf); /* fildll (%esp) */
 965         o(0x08c483); /* add $8, %esp */
 966     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 967                (VT_INT | VT_UNSIGNED)) {
 968         /* unsigned int to float/double/long double */
 969         o(0x6a); /* push $0 */
 970         g(0x00);
 971         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 972         o(0x242cdf); /* fildll (%esp) */
 973         o(0x08c483); /* add $8, %esp */
 974     } else {
 975         /* int to float/double/long double */
 976         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 977         o(0x2404db); /* fildl (%esp) */
 978         o(0x04c483); /* add $4, %esp */
 979     }
 980     vtop->r = TREG_ST0;
 981 }
 982
 983 /* convert fp to int 't' type */
 984 ST_FUNC void gen_cvt_ftoi(int t)
 985 {
 986  #ifndef COMMIT_4ad186c5ef61_IS_FIXED
 987     /* a good version but it takes a more time to execute */
 988     gv(RC_FLOAT);
 989     save_reg(TREG_EAX);
 990     save_reg(TREG_EDX);
 991     gen_static_call(TOK___tcc_cvt_ftol);
 992     vtop->r = TREG_EAX; /* mark reg as used */
 993     if (t == VT_LLONG)
 994         vtop->r2 = TREG_EDX;
 995  #else
 996     /* a new version with a bug: t2a = 44100312 */
 997     /*
 998     #include<stdio.h>
 999     int main() {
1000         int t1 = 176401255;
1001         float f = 0.25f;
1002         int t2a = (int)(t1 * f); // must be 44100313
1003         int t2b = (int)(t1 * (float)0.25f);
1004         printf("t2a=%d t2b=%d \n",t2a,t2b);
1005         return 0;
1006     }
1007     */
1008     int bt = vtop->type.t & VT_BTYPE;
1009     if (bt == VT_FLOAT)
1010         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1011     else if (bt == VT_LDOUBLE)
1012         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1013     else
1014         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1015     vswap();
1016     gfunc_call(1);
1017     vpushi(0);
1018     vtop->r = REG_IRET;
1019     vtop->r2 = REG_LRET;
1020  #endif
1021 }
1022
1023 /* convert from one floating point type to another */
1024 ST_FUNC void gen_cvt_ftof(int t)
1025 {
1026     /* all we have to do on i386 is to put the float in a register */
1027     gv(RC_FLOAT);
1028 }
1029
1030 /* computed goto support */
1031 ST_FUNC void ggoto(void)
1032 {
1033     gcall_or_jmp(1);
1034     vtop--;
1035 }
1036
1037 /* bound check support functions */
1038 #ifdef CONFIG_TCC_BCHECK
1039
1040 /* generate a bounded pointer addition */
1041 ST_FUNC void gen_bounded_ptr_add(void)
1042 {
1043     /* prepare fast i386 function call (args in eax and edx) */
1044     gv2(RC_EAX, RC_EDX);
1045     /* save all temporary registers */
1046     vtop -= 2;
1047     save_regs(0);
1048     /* do a fast function call */
1049     gen_static_call(TOK___bound_ptr_add);
1050     /* returned pointer is in eax */
1051     vtop++;
1052     vtop->r = TREG_EAX | VT_BOUNDED;
1053     /* address of bounding function call point */
1054     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1055 }
1056
1057 /* patch pointer addition in vtop so that pointer dereferencing is
1058    also tested */
1059 ST_FUNC void gen_bounded_ptr_deref(void)
1060 {
1061     addr_t func;
1062     addr_t size, align;
1063     Elf32_Rel *rel;
1064     Sym *sym;
1065
1066     size = 0;
1067     /* XXX: put that code in generic part of tcc */
1068     if (!is_float(vtop->type.t)) {
1069         if (vtop->r & VT_LVAL_BYTE)
1070             size = 1;
1071         else if (vtop->r & VT_LVAL_SHORT)
1072             size = 2;
1073     }
1074     if (!size)
1075         size = type_size(&vtop->type, &align);
1076     switch(size) {
1077     case  1: func = TOK___bound_ptr_indir1; break;
1078     case  2: func = TOK___bound_ptr_indir2; break;
1079     case  4: func = TOK___bound_ptr_indir4; break;
1080     case  8: func = TOK___bound_ptr_indir8; break;
1081     case 12: func = TOK___bound_ptr_indir12; break;
1082     case 16: func = TOK___bound_ptr_indir16; break;
1083     default:
1084         tcc_error("unhandled size when dereferencing bounded pointer");
1085         func = 0;
1086         break;
1087     }
1088
1089     /* patch relocation */
1090     /* XXX: find a better solution ? */
1091     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1092     sym = external_global_sym(func, &func_old_type, 0);
1093     if (!sym->c)
1094         put_extern_sym(sym, NULL, 0, 0);
1095     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1096 }
1097 #endif
1098
1099 /* Save the stack pointer onto the stack */
1100 ST_FUNC void gen_vla_sp_save(int addr) {
1101     /* mov %esp,addr(%ebp)*/
1102     o(0x89);
1103     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1104 }
1105
1106 /* Restore the SP from a location on the stack */
1107 ST_FUNC void gen_vla_sp_restore(int addr) {
1108     o(0x8b);
1109     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1110 }
1111
1112 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1113 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1114 #ifdef TCC_TARGET_PE
1115     /* alloca does more than just adjust %rsp on Windows */
1116     vpush_global_sym(&func_old_type, TOK_alloca);
1117     vswap(); /* Move alloca ref past allocation size */
1118     gfunc_call(1);
1119     vset(type, REG_IRET, 0);
1120 #else
1121     int r;
1122     r = gv(RC_INT); /* allocation size */
1123     /* sub r,%rsp */
1124     o(0x2b);
1125     o(0xe0 | r);
1126     /* We align to 16 bytes rather than align */
1127     /* and ~15, %esp */
1128     o(0xf0e483);
1129     /* mov %esp, r */
1130     o(0x89);
1131     o(0xe0 | r);
1132     vpop();
1133     vset(type, r, 0);
1134 #endif
1135 }
1136
1137 /* end of X86 code generator */
1138 /*************************************************************/
1139 #endif
1140 /*************************************************************/